Measuring 3D Shape when Light Misbehaves
Tuesday, June 14 at 11am in PT266
Several computer vision techniques assume that objects in a scene get illuminated only directly by the light source(s). For real world scenes, this is not true. Imagine a robot trying to navigate an underground cave, a surgical instrument inside the human body, a robotic arm sorting a heap of metallic machine parts, or a movie director wanting to image the face of an actor. In all these scenarios, light misbehaves - it gets reflected between different scene points, it scatters inside translucent materials (skin, plants) and volumetric media (smoke, murky water, liquids) and diffuses through semi-transparent objects like frosted glass. These intra-scene optical interactions, collectively called global light transport, often result in significant errors in the recovered scene properties. In order to develop computer vision systems which can perform reliably in real-world situations, it is imperative to account for these complex optical effects.
In this talk, I will discuss how global light transport influences the classical vision problem of shape recovery. I will talk about two different classes of shape recovery techniques (structured light scanning and shape from projector defocus) and show that they can be made nearly invariant to the effects of global light transport. I will demonstrate shape recovery results for a wide range of optically challenging scenes with complex geometry and material properties, such as shiny brushed metal, translucent marble and organic materials, and concave tubular objects.
Bio: Mohit Gupta is a post-doctoral research fellow working with Prof. Shree K. Nayar in the CAVE Lab, Columbia University. He received his B.Tech. in Computer Science from the Indian Institute of Technology, New Delhi in 2003, M.S. in Computer Science from the Stony Brook University in 2005 and Ph.D. in Robotics from the Robotics Institute, CMU. His research interests are in physics-based computer vision, global light transport, active illumination and computational imaging. He is passionate about building computer vision systems that can work in the real world. Find details about his research here.